Fixes #220.

Problem

cao launch (non-headless) attached to the tmux session the moment the API returned, before the provider's TUI had finished wiring its input handler. The attach-time pty resize would race with init and the TUI silently dropped keystrokes — most visible on agents with many MCP servers where init takes longer.

Root cause

cli/commands/launch.py had asymmetric wait semantics:

if not headless:
    subprocess.run(["tmux", "attach-session", "-t", terminal["session_name"]])  # no wait
elif message:
    wait_until_terminal_status(terminal["id"], {IDLE, COMPLETED}, timeout=120)   # proper wait

The headless-with-message path already had the wait. The non-headless path needed the same guard.

Fix

Add wait_until_terminal_status before tmux attach-session:

if not headless:
    ready = wait_until_terminal_status(terminal["id"], {IDLE, COMPLETED}, timeout=120)
    if not ready:
        click.echo(click.style("  Warning: ... attaching anyway; input may be unreliable ...", fg="yellow"))
    subprocess.run(["tmux", "attach-session", "-t", terminal["session_name"]])

The wait is advisory — on timeout we warn and still attach, because orphaning a slow-initializing session in tmux is worse than letting the user watch it come alive.

Why this is reliable now

PR #215 (false-IDLE during Initializing...) already landed, which means wait_until_terminal_status actually blocks through init instead of returning early on the idle-prompt placeholder that kiro-cli's TUI v2 paints before init completes.

Relationship to recent fixes

• kiro_cli: TUI v2 input unresponsive when session created headless (80x24 default pane size) #216 / fix(tmux): start panes at 220x50 to avoid kiro-cli SIGWINCH input dea… #218 (merged) — start pane at 220×50 so attach-time SIGWINCH is a shrink (kiro-cli 2.1.x handles shrinks correctly)
• fix(kiro_cli): detect TUI Initializing... to prevent false IDLE (#211) #215 (merged) — detect Initializing... to prevent false IDLE during init
• This PR (kiro_cli TUI: input unresponsive when tmux attach runs before kiro finishes initializing #220) — the missing piece: actually use the reliable wait on the non-headless attach path

All three address the same symptom (input dropped after launch) from different angles. They compose cleanly.

Changes

src/cli_agent_orchestrator/cli/commands/launch.py

• Insert wait_until_terminal_status before subprocess.run(["tmux", "attach-session", ...]) with a warn-and-proceed fallback on timeout.

test/cli/commands/test_launch.py

• New: test_launch_non_headless_waits_for_idle_before_attach — asserts the wait is called with the terminal id and runs before the attach.
• New: test_launch_non_headless_attaches_even_if_wait_times_out — asserts warn-and-proceed on timeout.
• Updated 4 existing non-headless tests to patch wait_until_terminal_status and include the terminal id in their mocked API response (previously the mock dict omitted id, which would now KeyError on the real code path).

Provider-agnostic

Although #220 was filed for kiro-cli, the wait applies to every provider. opencode_cli and others with slow TUI init also benefit.

Verification

• uv run pytest test/cli/commands/test_launch.py -q — 21/21 pass
• Full CI-filtered suite (--ignore=test/providers/test_q_cli_integration.py --ignore=test/providers/test_kiro_cli_integration.py --ignore=test/e2e -m "not e2e") — 1523 passed, 1 skipped
• black --check src/.../launch.py test/.../test_launch.py — clean
• isort --check-only src/.../launch.py test/.../test_launch.py — clean

Test plan

• Unit suite passes
• Lint + formatters clean
• Non-headless cao launch waits for IDLE before attach
• Non-headless launch still attaches if wait times out (with warning)
• Live: cao launch --agents developer --yolo on a profile with multiple MCP servers — keystrokes land reliably after attach